<html>
<title>
XHProf Documentation (Draft)
</title>
<body>
<h3>XHProf Documentation (Draft)</h3>
<b>Contents</b>
<ol>
<li><a href="#introduction">Introduction</a></li>
<li><a href="#overview">XHProf Overview</a></li>
<li><a href="#installation">Installing the XHProf extension</a></li>
<li><a href="#using_extension">Profiling using XHProf</a></li>
<li><a href="#ui_setup">Setting up the XHProf UI</a></li>
<li><a href="#production_notes">Notes on using XHProf in production</a></li>
<li><a href="#sampling_mode">Lightweight Sampling Mode</a>
<li><a href="#misc">Additional features</a></li>
<li><a href="#dependencies">Dependencies</a></li>
<li><a href="#credits">Acknowledgements</a></li>
</ol>
<ol>
<li><a name="introduction"><h2>Introduction</h2></a>
<p>XHProf is a hierarchical profiler for PHP. It reports
function-level call counts and <a href="#inclusive">inclusive</a> and
<a href="#exclusive">exclusive</a> metrics such as wall (elapsed)
time, CPU time and memory usage. A function's profile can be broken
down by callers or callees. The raw data collection component is
implemented in C as a PHP Zend extension called
<code><b>xhprof</b></code>. XHProf has a simple HTML based user
interface (written in PHP). The browser based UI for viewing profiler
results makes it easy to view results or to share results with peers.
A callgraph image view is also supported.
<p>XHProf reports can often be helpful in understanding the structure
of the code being executed. The hierarchical nature of the reports can
be used to determine, for example, what chain of calls led to a
particular function getting called.
<p>XHProf supports ability to compare two runs (a.k.a. "diff" reports)
or aggregate data from multiple runs. Diff and aggregate reports, much
like single run reports, offer "flat" as well as "hierarchical" views
of the profile.
<p>XHProf is a light-weight instrumentation based profiler. During the
data collection phase, it keeps track of call counts and inclusive
metrics for arcs in the dynamic callgraph of a program. It computes
exclusive metrics in the reporting/post processing phase. XHProf
handles recursive functions by detecting cycles in the callgraph at
data collection time itself and avoiding the cycles by giving unique
depth qualified names for the recursive invocations.
</p>
<p>XHProf's light-weight nature and aggregation capabilities make it
well suited for collecting "function-level" performance statistics
from production environments. [See <a
href="#production_notes">additional notes</a> for use in production.]
<ul>
<hr>
<p>XHProfLive (not part of the open source kit), for example, is a
system-wide performance monitoring system in use at Facebook that is
built on top of XHProf. XHProfLive continually gathers function-level
profiler data from production tier by running a sample of page
requests under XHProf. XHProfLive then aggregates the profile data
corresponding to individual requests by various dimensions such as
time, page type, and can help answer a variety of questions such as:
What is the function-level profile for a specific page? How expensive
is function "foo" across all pages, or on a specific page? What
functions regressed most in the last hour/day/week? What is the
historical trend for execution time of a page/function? and so on.
<hr>
</ul>
<p>Originally developed at Facebook, XHProf was open sourced in Mar, 2009.</p>
</ul>
<li><a name="overview"><h2>XHProf Overview</h2></a>
<p>XHProf provides:
<ul>
<li><b>Flat profile</b> (<a href="sample-flat-view.jpg" >screenshot</a>)
<p>Function-level summary information such as number of calls,
inclusive/exclusive wall time, memory usage, and CPU time.
<p><li><b>Hierarchical profile (Parent/Child View)</b>
(<a href="sample-parent-child-view.jpg" >screenshot</a>)
<p>For each function, it provides a breakdown of calls and times per
parent (caller) & child (callee), such as:
<ul>
<li> what functions call a particular function and how many times?
<li> what functions does a particular function call?
<li> The total time spent under a function when called from a particular parent.
</ul>
<p><li><b>Diff Reports</b>
<p>You may want to compare data from two XHProf runs for various
reasons-- to figure out what's causing a regression between one
version of the code base to another, to evaluate the performance
improvement of a code change you are making, and so on.
<p>A diff report takes two runs as input and provides both flat
function-level diff information, and hierarchical information
(breakdown of diff by parent/children functions) for each function.
<p>The "flat" view (<a href="sample-diff-report-flat-view.jpg"
>sample screenshot</a>) in the diff report points out the top
regressions & improvements.
<p>Clicking on functions in the "flat" view of the diff report, leads
to the "hierarchical" (or parent/child) diff view of a function (<a href="sample-diff-report-parent-child-view.jpg"
>sample screenshot</a>). We can get a
breakdown of the diff by parent/children functions.
<p><li><b>Callgraph View</b> (<a href="sample-callgraph-image.jpg"
>sample screenshot</a>)
<p>The profile data can also be viewed as a callgraph. The callgraph
view highlights the critical path of the program.
<p><li><b>Memory Profile</b>
<p>XHProf's memory profile mode helps track functions that
allocate lots of memory.
<p>It is worth clarifying that that XHProf doesn't strictly track each
allocation/free operation. Rather it uses a more simplistic
scheme. It tracks the increase/decrease in the amount of memory
allocated to PHP between each function's entry and exit. It also
tracks increase/decrease in the amount of <b>peak</b> memory allocated to
PHP for each function.
<li>XHProf tracks <code>include, include_once, require and
require_once</code> operations as if they were functions. The name of
the file being included is used to generate the name for these <a
href="#include_funcs">"fake" functions</a>.
</ul>
<a name="Terminology"></a><h2>Terminology</h2>
<ol>
<a name="inclusive"></a><li><b>Inclusive Time (or Subtree Time)</b>:
Includes time spent in the function as well as in descendant functions
called from a given function.
<a name="exclusive"></a><li><b>Exclusive Time/Self Time</b>: Measures
time spent in the function itself. Does not include time in descendant
functions.
<li><b>Wall Time</b>: a.k.a. Elapsed time or wall clock time.
<li><b>CPU Time</b>: CPU time in user space + CPU time in kernel space
</ol>
<a name="Naming_convention_for_special_functions"></a><h2>Naming convention for special functions</h2>
<ol>
<p><li><code><b>main()</b></code>: a fictitious function that is at the root of the call graph.
<a name="include_funcs"></a>
<p><li><code><b>load::<filename></b></code>
and <code><b>run_init::<filename></b></code>:
<p>XHProf tracks PHP <code>include/require</code> operations as
function calls.
<p>For example, an <b>include "lib/common.php";</b> operation will
result in two XHProf function entries:
<ul>
<li> <code><b>load::lib/common.php</b></code> - This represents the work done by the
interpreter to compile/load the file. [Note: If you are using a PHP
opcode cache like APC, then the compile only happens on a cache miss
in APC.]
<li> <code><b>run_init::lib/common.php</b></code> - This represents
initialization code executed at the file scope as a result of the
include operation.
</ul>
<p><li><code><b>foo@<n></b></code>: Implies that this is a
recursive invocation of <code>foo()</code>, where <code><n></code> represents
the recursion depth. The recursion may be direct (such as due to
<code>foo()</code> --> <code>foo()</code>), or indirect (such as
due to </code>foo()</code> --> <code>goo()</code> --> foo()).
</ol>
<a name="Limitations"></a><h2>Limitations</h2>
<p>True hierarchical profilers keep track of a full call stack at
every data gathering point, and are later able to answer questions
like: what was the cost of the 3rd invokation of foo()? or what was
the cost of bar() when the call stack looked like
a()->b()->bar()?
</p>
<p>XHProf keeps track of only 1-level of calling context and is
therefore only able to answer questions about a function looking
either 1-level up or 1-level down. It turns out that in practice this
is sufficient for most use cases.
</p>
<p>To make this more concrete, take for instance the following
example.
</p>
<pre>
Say you have:
1 call from a() --> c()
1 call from b() --> c()
50 calls from c() --> d()
</pre>
<p>While XHProf can tell you that d() was called from c() 50 times, it
cannot tell you how many of those calls were triggered due to a()
vs. b(). [We could speculate that perhaps 25 were due to a() and 25
due to b(), but that's not necessarily true.]
</p>
<p>In practice however, this isn't a very big limitation.
</p>
<li><a name="installation"><h2>Installing the XHProf Extension</h2></a>
<p> The extension lives in the "extension/" sub-directory.
<ul><hr>
<p><b>Note:</b> A windows port hasn't been implemented yet. We have
tested <code>xhprof</code> on <b>Linux/FreeBSD</b> so far.
<p>Version 0.9.2 and above of XHProf is also expected to work on <b>Mac
OS</b>. [We have tested on Mac OS 10.5.]
<p><b>Note:</b> XHProf uses the RDTSC instruction (time stamp counter)
to implement a really low overhead timer for elapsed time. So at the
moment <code>xhprof</code> only works on <b>x86</b> architecture.
Also, since RDTSC values may not be synchronized across CPUs,
<code>xhprof</code> binds the program to a single CPU during the
profiling period.
<p>XHProf's RDTSC based timer functionality doesn't work correctly if
<b>SpeedStep</b> technology is turned on. This technology is available on
some Intel processors. [Note: Mac desktops and laptops typically have
SpeedStep turned on by default. To use XHProf, you'll need to disable
SpeedStep.]
<hr></ul>
<p> The steps
below should work for Linux/Unix environments.
<pre>
% cd <xhprof_source_directory>/extension/
% phpize
% ./configure --with-php-config=<path to php-config>
% make
% make install
% make test
</pre>
<p><a name="ini_file"></a><b>php.ini file</b>: You can update your
php.ini file to automatically load your extension. Add the following
to your php.ini file.
<pre>
[xhprof]
extension=xhprof.so
;
; directory used by default implementation of the iXHProfRuns
; interface (namely, the XHProfRuns_Default class) for storing
; XHProf runs.
;
xhprof.output_dir=<directory_for_storing_xhprof_runs>
</pre>
<li><a name="using_extension"><h2>Profiling using XHProf</h2></a>
<p>Test generating raw profiler data using a sample test program like:
<ul>
<p><b>foo.php</b>
<pre>
<?php
function bar($x) {
if ($x > 0) {
bar($x - 1);
}
}
function foo() {
for ($idx = 0; $idx < 2; $idx++) {
bar($idx);
$x = strlen("abc");
}
}
// start profiling
<b>xhprof_enable();</b>
// run program
foo();
// stop profiler
<b>$xhprof_data = xhprof_disable();</b>
// display raw xhprof data for the profiler run
print_r($xhprof_data);
</pre>
</ul>
<p><b>Run the above test program:</b>
<pre>
% php -dextension=xhprof.so foo.php
</pre>
<p><b>You should get an output like:</b>
<pre>
Array
(
[foo==>bar] => Array
(
[ct] => 2 # 2 calls to bar() from foo()
[wt] => 27 # inclusive time in bar() when called from foo()
)
[foo==>strlen] => Array
(
[ct] => 2
[wt] => 2
)
[bar==>bar@1] => Array # a recursive call to bar()
(
[ct] => 1
[wt] => 2
)
[main()==>foo] => Array
(
[ct] => 1
[wt] => 74
)
[main()==>xhprof_disable] => Array
(
[ct] => 1
[wt] => 0
)
[main()] => Array # fake symbol representing root
(
[ct] => 1
[wt] => 83
)
)
</pre>
<p><b>Note:</b> The raw data only contains "inclusive" metrics. For
example, the wall time metric in the raw data represents inclusive
time in microsecs. Exclusive times for any function are computed
during the analysis/reporting phase.
<p><b>Note:</b> By default only call counts & elapsed time is profiled.
You can optionally also profile CPU time and/or memory usage. Replace,
<ul><pre>
<b>xhprof_enable();</b>
</pre></ul>
in the above program with, for example:
<ul><pre>
<b>xhprof_enable(XHPROF_FLAGS_CPU + XHPROF_FLAGS_MEMORY)</b>;
</pre></ul>
<p><b>You should now get an output like:</b>
<pre>
Array
(
[foo==>bar] => Array
(
[ct] => 2 # number of calls to bar() from foo()
[wt] => 37 # time in bar() when called from foo()
[cpu] => 0 # cpu time in bar() when called from foo()
[mu] => 2208 # change in PHP memory usage in bar() when called from foo()
[pmu] => 0 # change in PHP peak memory usage in bar() when called from foo()
)
[foo==>strlen] => Array
(
[ct] => 2
[wt] => 3
[cpu] => 0
[mu] => 624
[pmu] => 0
)
[bar==>bar@1] => Array
(
[ct] => 1
[wt] => 2
[cpu] => 0
[mu] => 856
[pmu] => 0
)
[main()==>foo] => Array
(
[ct] => 1
[wt] => 104
[cpu] => 0
[mu] => 4168
[pmu] => 0
)
[main()==>xhprof_disable] => Array
(
[ct] => 1
[wt] => 1
[cpu] => 0
[mu] => 344
[pmu] => 0
)
[main()] => Array
(
[ct] => 1
[wt] => 139
[cpu] => 0
[mu] => 5936
[pmu] => 0
)
)
</pre>
<p><b>Skipping builtin functions during profiling</b>
<p>By default PHP builtin functions (such as <code>strlen</code>) are
profiled. If you do not want to profile builtin functions (to either
reduce the overhead of profiling further or size of generated raw
data), you can use the <code><b>XHPROF_FLAGS_NO_BUILTINS</b></code>
flag as in for example:
<ul><pre>
// do not profile builtin functions
<b>xhprof_enable(XHPROF_FLAGS_NO_BUILTINS)</b>;
</pre></ul>
<p><b>Ignoring specific functions during profiling (0.9.2 or higher)</b>
<p>Starting with release 0.9.2 of xhprof, you can tell XHProf to
ignore a specified list of functions during profiling. This allows you
to ignore, for example, functions used for indirect function calls
such as <code>call_user_func</code> and
<code>call_user_func_array</code>. These intermediate functions
unnecessarily complicate the call hierarchy and make the XHProf
reports harder to interpret since they muddle the parent-child
relationship for functions called indirectly.
<p> To specify the list of functions to be ignored during profiling
use the 2nd (optional) argument to <code>xhprof_enable</code>.
For example,
<pre>
<ul><pre>
<b>
// elapsed time profiling; ignore call_user_func* during profiling
xhprof_enable(0,
array('ignored_functions' => array('call_user_func',
'call_user_func_array')));
</b>
or,
<b>
// elapsed time + memory profiling; ignore call_user_func* during profiling
xhprof_enable(XHPROF_FLAGS_MEMORY,
array('ignored_functions' => array('call_user_func',
'call_user_func_array')));
</b>
</pre></ul>
</li>
<li><a name="ui_setup"><h2>Setting up XHProf UI</h2></a>
<ol>
<li><b>PHP source structure</b>
<p>The XHProf UI is implemented in PHP. The code resides in two
subdirectories, <code>xhprof_html/</code> and <code>xhprof_lib/</code>.
<p>The <code>xhprof_html</code> directory contains the 3 top-level PHP pages.
<ul>
<li><code>index.php</code>: For viewing a single run or diff report.
<li><code>callgraph.php</code>: For viewing a callgraph of a XHProf run as an image.
<li><code>typeahead.php</code>: Used implicitly for the function typeahead form
on a XHProf report.
</ul>
<p>The <code>xhprof_lib</code> directory contains supporting code for
display as well as analysis (computing flat profile info, computing
diffs, aggregating data from multiple runs, etc.).
<li><p><b>Web server config: </b> You'll need to make sure that the
<code>xhprof_html/</code> directory is accessible from your web server, and that
your web server is setup to serve PHP scripts.
<li><p><b>Managing XHProf Runs</b>
<p>Clients have flexibility in how they save the XHProf raw data
obtained from an XHProf run. The XHProf UI layer exposes an interface
iXHProfRuns (see xhprof_lib/utils/xhprof_runs.php) that clients can
implement. This allows the clients to tell the UI layer how to fetch
the data corresponding to a XHProf run.
<p>The XHProf UI libaries come with a default file based
implementation of the iXHProfRuns interface, namely
"XHProfRuns_Default" (also in xhprof_lib/utils/xhprof_runs.php).
This default implementation stores runs in the directory specified by
<a href="#ini_file"><b>xhprof.output_dir</b></a> INI parameter.
<p>A XHProf run must be uniquely identified by a namespace and a run
id.
<p><b>a) Saving XHProf data persistently</b>:
<p>Assuming you are using the default implementation
<code><b>XHProfRuns_Default</b></code> of the
<code><b>iXHProfRuns</b></code> interface, a typical XHProf run
followed by the save step might look something like:
<pre>
// start profiling
xhprof_enable();
// run program
....
// stop profiler
$xhprof_data = xhprof_disable();
//
// Saving the XHProf run
// using the default implementation of iXHProfRuns.
//
include_once $XHPROF_ROOT . "/xhprof_lib/utils/xhprof_lib.php";
include_once $XHPROF_ROOT . "/xhprof_lib/utils/xhprof_runs.php";
$xhprof_runs = new <b>XHProfRuns_Default()</b>;
// Save the run under a namespace "xhprof_foo".
//
// **NOTE**:
// By default save_run() will automatically generate a unique
// run id for you. [You can override that behavior by passing
// a run id (optional arg) to the save_run() method instead.]
//
<b>$run_id = $xhprof_runs->save_run($xhprof_data, "xhprof_foo");</b>
echo "---------------\n".
"Assuming you have set up the http based UI for \n".
"XHProf at some address, you can view run at \n".
"http://<xhprof-ui-address>/index.php?run=$run_id&source=xhprof_foo\n".
"---------------\n";
</pre>
<p>The above should save the run as a file in the directory specified
by the <code><b>xhprof.output_dir</b></code> INI parameter. The file's
name might be something like
<b><code>49bafaa3a3f66.xhprof_foo</code></b>; the two parts being the
run id ("49bafaa3a3f66") and the namespace ("xhprof_foo"). [If you
want to create/assign run ids yourself (such as a database sequence
number, or a timestamp), you can explicitly pass in the run id to the
<code>save_run</code> method.
<p><b>b) Using your own implementation of iXHProfRuns</b>
<p> If you decide you want your XHProf runs to be stored differently
(either in a compressed format, in an alternate place such as DB,
etc.) database, you'll need to implement a class that implements the
iXHProfRuns() interface.
<p> You'll also need to modify the 3 main PHP entry pages (index.php,
callgraph.php, typeahead.php) in the "xhprof_html/" directory to use
the new class instead of the default class <code>XHProfRuns_Default</code>.
Change this line in the 3 files.
<pre>
$xhprof_runs_impl = new XHProfRuns_Default();
</pre>
<p>You'll also need to "include" the file that implements your class in
the above files.
<li><p><b>Accessing runs from UI</b>
<p><b>a) Viewing a Single Run Report</b>
<p>To view the report for run id say <run_id> and namespace
<namespace> use a URL of the form:
<p><code>
http://<xhprof-ui-address>/index.php?run=<run_id>&source=<namespace>
</code>
<p>For example,
<p><code>
http://<xhprof-ui-address>/index.php?run=49bafaa3a3f66&source=xhprof_foo
</code>
<p><b>b) Viewing a Diff Report</b>
<p>To view the report for run ids say <run_id1> and
<run_id2> in namespace <namespace> use a URL of the form:
<p><code>
http://<xhprof-ui-address>/index.php?<b>run1=<run_id1>&run2=<run_id2></b>&source=<namespace>
</code>
<p><b>c) Aggregate Report</b>
<p>You can also specify a set of run ids for which you want an aggregated view/report.
<p>Say you have three XHProf runs with ids 1, 2 & 3 in namespace
"benchmark". To view an aggregate report of these runs:
<ul><p><code>
http://<xhprof-ui-address>/index.php?<b>run=1,2,3</b>&source=benchmark
</code></p></ul>
<p><b>Weighted aggregations</b>: Further suppose that the above three runs
correspond to three types of programs p1.php, p2.php and p3.php that
typically occur in a mix of 20%, 30%, 50% respectively. To view an
aggregate report that corresponds to a weighted average of these runs
using:
<ul><p><code>
http://<xhprof-ui-address>/index.php?<b>run=1,2,3&wts=20,30,50</b>&source=benchmark
</code></p></ul>
</ol>
<li><a name="production_notes"><h2>Notes on using XHProf in production</h2></a>
<p>Some observations/guidelines. Your mileage may vary:
<ul>
<li>CPU timer (getrusage) on Linux has high overheads. It is also
coarse grained (millisec accuracy rather than microsec level) to be
useful at function level. Therefore, the skew in reported numbers
when using XHPROF_FLAGS_CPU mode tends to be higher.
<p>We recommend using elapsed time + memory profiling mode in
production. [Note: The additional overhead of memory profiling
mode is really low.]
<p><ul><pre><b>
// elapsed time profiling (default) + memory profiling
xhprof_enable(XHPROF_FLAGS_MEMORY);
</b></pre>
</ul></p>
<li>Profiling a random sample of pages/requests works well in capturing
data that is representative of your production workload.
<p>To profile say 1/10000 of your requests, instrument the beginning of
your request processing with something along the lines of:
<p><ul><pre><code>
if (mt_rand(1, 10000) == 1) {
xhprof_enable(XHPROF_FLAGS_MEMORY);
$xhprof_on = true;
}
</code></pre></ul></p>
<p>At the end of the request (or in a request shutdown function), you might
then do something like:
<p><ul><pre><code>
if ($xhprof_on) {
// stop profiler
$xhprof_data = xhprof_disable();
// save $xhprof_data somewhere (say a central DB)
...
}
</code></pre></ul></p>
<p> You can then rollup/aggregate these individual profiles by time
(e.g., 5 minutely/hourly/daily basis), page/request type,or other
dimensions using <a href="#xhprof_aggregate_runs"><code><b>xhprof_aggregate_runs()</b></code></a>.
</ul>
<li><a name="sampling_mode"><h2>Lightweight Sampling Mode</h2></a>
<p>The xhprof extension also provides a very light weight <b>sampling
mode</b>. The sampling interval is 0.1 secs. Samples record the full
function call stack. The sampling mode can be useful if an extremely
low overhead means of doing performance monitoring and diagnostics is
desired.
<p> The relevant functions exposed by the extension for using the
sampling mode are <code><b>xhprof_sample_enable()</b></code> and
<code><b>xhprof_sample_disable()</b></code>.
<p>[<b>TBD</b>: more detailed documentation on sampling mode.]
<li><a name="misc"><h2>Additional Features</h2></a></li>
<p>The <code><b>xhprof_lib/utils/xhprof_lib.php</b></code> file contains
additional library functions that can be used for manipulating/
aggregating XHProf runs.
<p>For example:
<ul>
<a name="xhprof_aggregate_runs"></a>
<p><li><code><b>xhprof_aggregate_runs()</b></code>:
can be used to aggregate multiple XHProf runs into a single run. This
can be helpful for building a system-wide "function-level" performance
monitoring tool using XHProf. [For example, you might to roll up
XHProf runs sampled from production periodically to generate hourly,
daily, reports.]
<p><li><code><b>xhprof_prune_run()</b></code>: Aggregating large number of
XHProf runs (especially if they correspond to different types of
programs) can result in the callgraph size becoming too large. You can
use <code>xhprof_prune_run</code> function to prune the callgraph data
by editing out subtrees that account for a very small portion of the
total time.
</ul>
<ol>
</ol>
<li><a name="dependencies"><h2>Dependencies</h2></a></li>
<ul>
<li><b>JQuery Javascript</b>: For tooltips and function name
typeahead, we make use of JQuery's javascript libraries. JQuery is
available under both a MIT and GPL licencse
(http://docs.jquery.com/Licensing). The relevant JQuery code, used by
XHProf, is in the <code>xhprof_html/jquery</code> subdirectory.
<li><b>dot (image generation utility):</b> The callgraph image
visualization ([View Callgraph]) feature relies on the presence of
Graphviz "dot" utility in your path. "dot" is a utility to
draw/generate an image for a directed graph.
</ul>
<li><a name="credits"><h2>Acknowledgements</h2></a>
<p>The HTML-based navigational interface for browsing profiler results
is inspired by that of a similar tool that exists for Oracle's stored
procedure language, PL/SQL. But that's where the similarity ends; the
internals of the profiler itself are quite different.
</li>
</ol>
</body>
</html>
